X-Moto's batches copied at every draw (2026-10-18)

Moved on 2026-10-26 from docs/HOTPATHS.md (“X-Moto’s client arrays” and the “Measured on the PS3” rows), where Mesa’s side of client vertex arrays stays.

Symptom

The first profiling batch (2026-10-18, Mesa ps3.299) put 5.5% of X-Moto’s CPU (profile 013240, the credits replay at 154 fps) in u_upload_data, Mesa’s copy of client vertex arrays: memcpy < u_upload_data 5.3%, the biggest single X-Moto item in the graphics stack.

Evidence

All of it is xmoto’s own batches (patch 0001 batches DrawLib’s primitives per state run), drawn from a client array that Mesa copies at every draw. 3.9 of the 5.5% is the minimap: renderMiniMap > drawCircle, a circle of 8 segments per sprite entity in the minimap’s box (933 a frame in the credits replay, ~17000 batch vertices). The copy goes into GART, which ps3rsx maps cacheable (ordinary RAM, IOMMU-coherent).

Fix

xmoto 0004 XMOTO_GL_STREAM=1 (0, the default: as before; recipe variant stream). The batch is written straight into a 2 MiB GL_STREAM_DRAW buffer, mapped with glMapBufferRange (write, unsynchronized, invalidated range, explicit flush) from a batch’s first vertex to its draw, which sources the buffer at the batch’s offset: the vertices are written once, where the GPU reads them. When the next batch may not fit, the buffer gets new storage (glBufferData NULL), so nothing the GPU may still read is written again; nothing else reaches the GL while it is mapped (every other call draws what is pending first). Same vertices, order and draws. In nv30 an unsynchronized map of a GART buffer is the buffer’s own mapping, no staging. The llvmpipe check gamedb/games/xmoto/packages/xmoto/test/glbatch-test.sh draws a random scene through the layer for 400 frames (about 950 batches and one buffer renewal a frame): the same pixels with and without, no GL errors.

On the PS3, one run each the first time, then x3 in batch 16:

Numbers: batch 2026-10-18-mesa-416-hot-spots (Mesa ps3.416~dev hot spots (staging pool, query wait), cube2 355 oqlag, xmoto 362 stream, ClassiCube chunks1 (kick), 2026-10-18; medians over the runs of each game and variant).

Numbers: batch 2026-10-18-batch16 (batch 16: game prefetch A/B, idle fix on/off (433), ClassiCube chunks1, X-Moto stream, Luanti lean parts, 2026-10-18; medians over the runs of each game and variant).

A small fps gain (163.1 against 160.4, then 162.1 against 160.7: +1%); the “fewer hitches” of the first run (longest gameplay frame 146 against 404 ms) did not repeat (380 against 43 ms). It stays off; it is one of the switch-gated changes in docs/TESTS.md’s keep-or-drop re-test. Look for missing or flickering minimap dots and sprites on the webcam when it runs.